Abstract: Multi-camera fusion is a critical technology for achieving comprehensive perception of coal mine heading faces, and extrinsic calibration of multi-camera systems is a fundamental prerequisite for multi-camera fusion. However, due to the confined space and equipment occlusion in heading faces, the fields of view of multi-cameras mounted on heading equipment often have minimal or even no overlap, which poses significant challenges to multi-camera extrinsic calibration. To address this issue, an extrinsic calibration method fusing multi-target and planar constraints is proposed for on-board multi-cameras of heading equipment. Firstly, an initialization method for multi-camera extrinsic parameters based on a multi-target factor graph is designed. By using calibration targets as a bridge, a multi-target factor graph model is constructed, which converts the observations of the moving auxiliary camera on the targets and the prior geometric relationships between targets into factor constraints. This solves the problem that the on-board multi-cameras of heading equipment cannot establish correspondences due to the lack of common-view features, and achieves the effective acquisition of initial values of on-board multi-camera extrinsic parameters. Secondly, a global optimization method for multi-camera extrinsic parameters based on multi-planar constraints is proposed. By fusing the consistency constraints between target poses and point cloud planes, the multi-camera extrinsic parameters are globally optimized to further improve the accuracy of extrinsic calibration. Finally, abundant simulation experiments and real underground coal mine experiments are conducted to verify the effectiveness of the method under different camera arrangements, overlapping field of view ratios, and interference conditions. The experimental results show that the proposed calibration method significantly outperforms the traditional stereo calibration method in terms of accuracy when the field of view overlap is insufficient and common-view features are lacking. Specifically, the proposed method achieves a translation distance error of 0.0229 m, an angular error of 4.60°, and a relative distance error of 0.84%, while the traditional stereo calibration method yields a translation distance error of 0.6188 m, an angular error of 17.07°, and a relative distance error of 23.01%.